Auto-segmentation editing burden in head and neck planning

This is an OncoWiz educational overview of a research area, written for clinicians. It summarises the shape of the evidence and the questions worth asking of it. It is not a summary of any single study, and it reports no individual trial’s results.

Automatic contouring is the most clinically embedded use of AI in oncology today. In radiotherapy planning it is routinely deployed, and it is deployed in the honest configuration: the model proposes contours and a clinician edits them. That makes editing burden, not raw accuracy, the metric that decides whether it helps.

Why geometric overlap is the wrong headline

Segmentation quality is usually reported as overlap with a reference contour. That measure is dominated by large structures, so a system can post excellent aggregate scores while requiring substantial correction on exactly the small structures that take the longest to draw.

In head and neck planning this mismatch is at its sharpest. The region contains many small, low-contrast organs at risk sitting close to target volumes, and a millimetre of contour error there carries more dosimetric consequence than a larger error elsewhere.

Time saved versus time spent correcting

The relevant measurement is end-to-end: total clinician time from the start of contouring to an approved plan, with and without automation, on comparable cases. Reported gains are usually real but smaller than raw contouring-time comparisons suggest, because the corrections that remain are concentrated in the hardest structures.

• Large, high-contrast structures are where automation saves the most and needs the least editing.
• Small serial organs and structures defined partly by anatomical convention rather than image contrast need the most correction.
• Post-surgical anatomy, dental artefact and unusual presentations degrade output disproportionately.
• Editing a poor contour can take longer than drawing one from scratch, which is why a per-structure view matters.

The consistency argument

Time is not the only benefit, and may not be the main one. Automatic contouring reduces inter-observer variability, which matters for plan quality, for multi-centre trial compliance and for auditability. A department may reasonably adopt it for consistency even where the time saving is modest.

Before and after deployment

• Measure editing time per structure on your own cases, not overlap on the vendor’s.
• Confirm the model was trained on protocols and contouring conventions close to your own.
• Keep clinician approval mandatory, and make the editing step visible in the record.
• Watch for automation bias: approving contours without real inspection is the failure mode this workflow creates.
• Re-audit after any change to scanner, protocol or contouring guideline.